We see connection timeouts and other access issues when run exactly on
the hour, either due to load on their servers from similar cron jobs,
or from operational processes of theirs.
Move to on the :17s to avoid these access issues.
Increasing worker_connections has a memory cost, unlike the rest of
the changes in 1c76036c61d8; setting it to 1 million caused nginx to
consume several GB of memory.
Reduce the default down to 10k, and allow deploys to configure it up
if necessary. `worker_rlimit_nofile` is left at 1M, since it has no
impact on memory consumption.
There is no reason that the base node access method should be run
under supervisor, which exists primarily to give access to the `zulip`
user to restart its managed services. This access is unnecessary for
Teleport, and also causes unwanted restarts of Teleport services when
the `supervisor` base configuration changes. Additionally,
supervisor does not support the in-place upgrade process that Teleport
uses, as it replaces its core process with a new one.
Switch to installing a systemd configuration file (as generated by
`teleport install systemd`) for each part of Teleport, customized to
pass a `--config` path. As such, we explicitly disable the `teleport`
service provided by the package.
The supervisor process is shut down by dint of no longer installing
the file, which purges it from the managed directory, and reloads
Supervisor to pick up the removed service.
Zulip already has integrations for server-side Sentry integration;
however, it has historically used the Zulip-specific `blueslip`
library for monitoring browser-side errors. However, the latter sends
errors to email, as well optionally to an internal `#errors` stream.
While this is sufficient for low volumes of users, and useful in that
it does not rely on outside services, at higher volumes it is very
difficult to do any analysis or filtering of the errors. Client-side
errors are exceptionally noisy, with many false positives due to
browser extensions or similar, so determining real real errors from a
stream of un-grouped emails or messages in a stream is quite
difficult.
Add a client-side Javascript sentry integration. To provide useful
backtraces, this requires extending the pre-deploy hooks to upload the
source-maps to Sentry. Additional keys are added to the non-public
API of `page_params` to control the DSN, realm identifier, and sample
rates.
The current threshold of 40k descriptors was set in 2016, chosen to be
"at least 40x our current scale." At present, that only provides a
50% safety margin. Increase to 1 million to provide the same 40x
buffer as previously.
The highest value currently allowed by the kernels in
production (linux 5.3.0) is 1048576. This is set as the hard limit.
The 1 million limit is likely far above what the system can handle for
other reasons (memory, cpu, etc). While this removes a potential
safeguard on overload due to too many connections, due to the longpoll
architecture we would generally prefer to service more connections at
lower quality (due to CPU limitations) rather than randomly reject
additional connections.
Relevant prior commits:
- 836f313e69
- f2f97dd335
- ec23996538
- 8806ec698a
- e4fce10f46
After reflecting a bit on the last commit, I think it's substantially
easier to understand what's happening for these two tasks to be
defined in the same file, because we want the timing to be different
to avoid potential races.
5db55c38dc switched from `ensure => present` to the more specific
`ensure => directory` on the premise that tarballs would result in
more than one file being copied out of them. However, we only extract
a single file from the wal-g tarball, and install it at the output
path. The new rule attempts to replace it with an empty directory
after extraction.
Switch back to `ensure => present` for the tarball codepath.
These hooks are run immediately around the critical section of the
upgrade. If the upgrade fails for preparatory reasons, the pre-deploy
hook may not be run; if it fails during the upgrade, the post-deploy
hook will not be run. Hooks are called from the CWD of the new
deploy, with arguments of the old version and the new version. If
they exit with non-0 exit code, the deploy aborts.
Similar to the previous commit, Django was responsible for setting the
Content-Disposition based on the filename, whereas the Content-Type
was set by nginx based on the filename. This difference is not
exploitable, as even if they somehow disagreed with Django's expected
Content-Type, nginx will only ever respond with Content-Types found in
`uploads.types` -- none of which are unsafe for user-supplied content.
However, for consistency, have Django provide both Content-Type and
Content-Disposition headers.
The Content-Type of user-provided uploads was provided by the browser
at initial upload time, and stored in S3; however, 04cf68b45e
switched to determining the Content-Disposition merely from the
filename. This makes uploads vulnerable to a stored XSS, wherein a
file uploaded with a content-type of `text/html` and an extension of
`.png` would be served to browsers as `Content-Disposition: inline`,
which is unsafe.
The `Content-Security-Policy` headers in the previous commit mitigate
this, but only for browsers which support them.
Revert parts of 04cf68b45e, specifically by allowing S3 to provide
the Content-Disposition header, and using the
`ResponseContentDisposition` argument when necessary to override it to
`attachment`. Because we expect S3 responses to vary based on this
argument, we include it in the cache key; since the query parameter
has dashes in it, we can't use use the helper `$arg_` variables, and
must parse it from the query parameters manually.
Adding the disposition may decrease the cache hit rate somewhat, but
downloads are infrequent enough that it is unlikely to have a
noticeable effect. We take care to not adjust the cache key for
requests which do not specify the disposition.
This was missed in 04cf68b45ebb5c03247a0d6453e35ffc175d55da; as this
content is fundamentally untrusted, it must be served with
`Content-Security-Policy` headers in order to be safe. These headers
were not provided previously for S3 content because it was served from
the S3 domain.
This mitigates content served from Zulip which could be a stored XSS,
but only in browsers which support Content-Security-Policy headers;
see subsequent commit for the complete solution.
In nginx, `location` blocks operate on the _decoded_ URI[^1]:
> The matching is performed against a normalized URI, after decoding
> the text encoded in the “%XX” form
This means that if a user-uploaded file contains characters that are
not URI-safe, the browser encodes them in UTF-8 and then URI-encodes
them -- and nginx decodes them and reassembles the original character
before running the `location ~ ^/...` match. This means that the `$2`
_is not URI-encoded_ and _may contain non-ASCII characters.
When `proxy_pass` is passed a value containing one or more variables,
it does no encoding on that expanded value, assuming that the bytes
are exactly as they should be passed to the upstream. This means that
directly calling `proxy_pass https://$1/$2` would result in sending
high-bit characters to the S3 upstream, which would rightly balk.
However, a longstanding bug in nginx's `set` directive[^2] means that
the following line:
```nginx
set $download_url https://$1/$2;
```
...results in nginx accidentally URI-encoding $1 and $2 when they are
inserted, resulting in a `$download_url` which is suitable to pass to
`proxy_pass`. This bug is only present with numeric capture
variables, not named captures; this is particularly relevant because
numeric captures are easily overridden by additional regexes
elsewhere, as subsequent commits will add.
Fixing this is complicated; nginx does not supply any way to escape
values[^3], besides a third-party module[^4] which is an undue
complication to begin using. The only variable which nginx exposes
which is _not_ un-escaped already is `$request_uri`, which contains
the very original URL sent by the browser -- and thus can't respect
any work done in Django to generate the `X-Accel-Redirect` (e.g., for
`/user_uploads/temporary/` URLs). We also cannot pass these URLs to
nginx via query-parameters, since `$arg_foo` values are not
URI-decoded by nginx, there is no function to do so[^3], and the
values must be URI-encoded because they themselves are URLs with query
parameters.
Extra-URI-encode the path that we pass to the `X-Accel-Redirect`
location, for S3 redirects. We rely on the `location` block
un-escaping that layer, leaving `$s3_hostname` and `$s3_path` as they
were intended in Django.
This works around the nginx bug, with no behaviour change.
[^1]: http://nginx.org/en/docs/http/ngx_http_core_module.html#location
[^2]: https://trac.nginx.org/nginx/ticket/348
[^3]: https://trac.nginx.org/nginx/ticket/52
[^4]: https://github.com/openresty/set-misc-nginx-module#set_escape_uri
Black 23 enforces some slightly more specific rules about empty line
counts and redundant parenthesis removal, but the result is still
compatible with Black 22.
(This does not actually upgrade our Python environment to Black 23
yet.)
Signed-off-by: Anders Kaseorg <anders@zulip.com>
The `postfix.mailname` setting in `/etc/zulip.conf` was previously
only used for incoming mail, to identify in Postfix configuration
which messages were "local."
Also set `/etc/mailname`, which is used by Postfix to set how it
identifies to other hosts when sending outgoing email.
Co-authored-by: Alex Vandiver <alexmv@zulip.com>
Puppet _always_ sets the `+x` bit on directories if they have the `r`
bit set for that slot[^1]:
> When specifying numeric permissions for directories, Puppet sets the
> search permission wherever the read permission is set.
As such, for instance, `0640` is actually applied as `0750`.
Fix what we "want" to match what puppet is applying, by adding the `x`
bit. In none of these cases did we actually intend the directory to
not be executable.
[1] https://www.puppet.com/docs/puppet/5.5/types/file.html#file-attribute-mode
This was last really used in d7a3570c7e, in 2013, when it was
`/home/humbug/logs`.
Repoint the one obscure piece of tooling that writes there, and remove
the places that created it.
Zulip runs puppet manually, using the command-line tool; it does not
make use of the `puppet` service which, by default, attempts to
contact a host named `puppet` every two minutes to get a manifest to
apply. These attempts can generate log spam and user confusion.
Disable and stop the `puppet` service via puppet.
When file uploads are stored in S3, this means that Zulip serves as a
302 to S3. Because browsers do not cache redirects, this means that
no image contents can be cached -- and upon every page load or reload,
every recently-posted image must be re-fetched. This incurs extra
load on the Zulip server, as well as potentially excessive bandwidth
usage from S3, and on the client's connection.
Switch to fetching the content from S3 in nginx, and serving the
content from nginx. These have `Cache-control: private, immutable`
headers set on the response, allowing browsers to cache them locally.
Because nginx fetching from S3 can be slow, and requests for uploads
will generally be bunched around when a message containing them are
first posted, we instruct nginx to cache the contents locally. This
is safe because uploaded file contents are immutable; access control
is still mediated by Django. The nginx cache key is the URL without
query parameters, as those parameters include a time-limited signed
authentication parameter which lets nginx fetch the non-public file.
This adds a number of nginx-level configuration parameters to control
the caching which nginx performs, including the amount of in-memory
index for he cache, the maximum storage of the cache on disk, and how
long data is retained in the cache. The currently-chosen figures are
reasonable for small to medium deployments.
The most notable effect of this change is in allowing browsers to
cache uploaded image content; however, while there will be many fewer
requests, it also has an improvement on request latency. The
following tests were done with a non-AWS client in SFO, a server and
S3 storage in us-east-1, and with 100 requests after 10 requests of
warm-up (to fill the nginx cache). The mean and standard deviation
are shown.
| | Redirect to S3 | Caching proxy, hot | Caching proxy, cold |
| ----------------- | ------------------- | ------------------- | ------------------- |
| Time in Django | 263.0 ms ± 28.3 ms | 258.0 ms ± 12.3 ms | 258.0 ms ± 12.3 ms |
| Small file (842b) | 586.1 ms ± 21.1 ms | 266.1 ms ± 67.4 ms | 288.6 ms ± 17.7 ms |
| Large file (660k) | 959.6 ms ± 137.9 ms | 609.5 ms ± 13.0 ms | 648.1 ms ± 43.2 ms |
The hot-cache performance is faster for both large and small files,
since it saves the client the time having to make a second request to
a separate host. This performance improvement remains at least 100ms
even if the client is on the same coast as the server.
Cold nginx caches are only slightly slower than hot caches, because
VPC access to S3 endpoints is extremely fast (assuming it is in the
same region as the host), and nginx can pool connections to S3 and
reuse them.
However, all of the 648ms taken to serve a cold-cache large file is
occupied in nginx, as opposed to the only 263ms which was spent in
nginx when using redirects to S3. This means that to overall spend
less time responding to uploaded-file requests in nginx, clients will
need to find files in their local cache, and skip making an
uploaded-file request, at least 60% of the time. Modeling shows a
reduction in the number of client requests by about 70% - 80%.
The `Content-Disposition` header logic can now also be entirely shared
with the local-file codepath, as can the `url_only` path used by
mobile clients. While we could provide the direct-to-S3 temporary
signed URL to mobile clients, we choose to provide the
served-from-Zulip signed URL, to better control caching headers on it,
and greater consistency. In doing so, we adjust the salt used for the
URL; since these URLs are only valid for 60s, the effect of this salt
change is minimal.
Moving `/user_avatars/` to being served partially through Django
removes the need for the `no_serve_uploads` nginx reconfiguring when
switching between S3 and local backends. This is important because a
subsequent commit will move S3 attachments to being served through
nginx, which would make `no_serve_uploads` entirely nonsensical of a
name.
Serve the files through Django, with an offload for the actual image
response to an internal nginx route. In development, serve the files
directly in Django.
We do _not_ mark the contents as immutable for caching purposes, since
the path for avatar images is hashed only by their user-id and a salt,
and as such are reused when a user's avatar is updated.
The `django-sendfile2` module unfortunately only supports a single
`SENDFILE` root path -- an invariant which subsequent commits need to
break. Especially as Zulip only runs with a single webserver, and
thus sendfile backend, the functionality is simple to inline.
It is worth noting that the following headers from the initial Django
response are _preserved_, if present, and sent unmodified to the
client; all other headers are overridden by those supplied by the
internal redirect[^1]:
- Content-Type
- Content-Disposition
- Accept-Ranges
- Set-Cookie
- Cache-Control
- Expires
As such, we explicitly unset the Content-type header to allow nginx to
set it from the static file, but set Content-Disposition and
Cache-Control as we want them to be.
[^1]: https://www.nginx.com/resources/wiki/start/topics/examples/xsendfile/
As uploads are a feature of the application, not of a generic nginx
deployment, move them into the `zulip::app_frontend_base` class. This
is purely for organizational clarity -- we do not support deployments
with has `zulip::nginx` but not `zulip::app_frontend_base`.
‘exit’ is pulled in for the interactive interpreter as a side effect
of the site module; this can be disabled with python -S and shouldn’t
be relied on.
Also, use the NoReturn type where appropriate.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Starting with wal-g 2.0.1, they provide `aarch64` assets[^1].
Effectively revert d7b59c86ce, and use
the pre-built binary for `aarch64` rather than spend a bunch of space
and time having to build it from source.
[^1]: https://github.com/wal-g/wal-g/releases/tag/v2.0.1
A number of autossh connections are already left open for
port-forwarding Munin ports; autossh starts the connections and
ensures that they are automatically restarted if they are severed.
However, this represents a missed opportunity. Nagios's monitoring
uses a large number of SSH connections to the remote hosts to run
commands on them; each of these connections requires doing a complete
SSH handshake and authentication, which can have non-trivial network
latency, particularly for hosts which may be located far away, in a
network topology sense (up to 1s for a no-op command!).
Use OpenSSH's ability to multiplex multiple connections over a single
socket, to reuse the already-established connection. We leave an
explicit `ControlMaster no` in the general configuration, and not
`auto`, as we do not wish any of the short-lived Nagios connections to
get promoted to being a control socket if the autossh is not running
for some reason.
We enable protocol-level keepalives, to give a better chance of the
socket being kept open.
These hosts were excluded from `zulipconf_nagios_hosts` in
8cff27f67d, because it was replicating the previously hard-coded
behaviour exactly. That behaviour was an accident of history, in that
4fbe201187 and before had simply not monitored hosts of this class.
There is no reason to not add SSH tunnels and munin monitoring for
these hosts; stop skipping them.
Since Django factors request.is_secure() into its CSRF check, we need
this to tell it to consider requests forwarded from nginx to Tornado
as secure.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
One should now be able to configure a regex by appending _regex to the
port number:
[tornado_sharding]
9802_regex = ^[l-p].*\.zulipchat\.com$
Signed-off-by: Anders Kaseorg <anders@zulip.com>
https://nginx.org/en/docs/http/ngx_http_map_module.html
Since Puppet doesn’t manage the contents of nginx_sharding.conf after
its initial creation, it needs to be renamed so we can give it
different default contents.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Some legitimate requests in Zulip can take more than 20s to be
processed, and we don't have a current problem where having a 20s
limit here is preventing a problem.
The `needrestart` tool added in 22.04 is useful in terms of listing
which services may need to be restarted to pick up updated libraries.
However, it prompts about the current state of services needing
restart for *every* subsequent `apt-get upgrade`, and defaulting core
services to restarting requires carefully manually excluding them
every time, at risk of causing an unscheduled outage.
Build a list of default-off services based on the list in
unattended-upgrades.
The default value in uwsgi is 4k; receiving more than this amount from
nginx leads to a 502 response (though, happily, the backend uwsgi does not
terminate).
ab18dbfde5 originally increased it from the unstated uwsgi default
of 4096, to 8192; b1da797955 made it configurable, in order to allow
requests from clients with many cookies, without causing 502's[1].
nginx defaults to a limitation of 1k, with 4 additional 8k header
lines allowed[2]; any request larger than that returns a response of
`400 Request Header Or Cookie Too Large`. The largest header size
theoretically possible from nginx, by default, is thus 33k, though
that would require packing four separate headers to exactly 8k each.
Remove the gap between nginx's limit and uwsgi's, which could trigger
502s, by removing the uwsgi configurability, and setting a 64k size in
uwsgi (the max allowable), which is larger than nginx's default limit.
uWSGI's documentation of `buffer-size` ([3], [4]) also notes that "It
is a security measure too, so adapt to your app needs instead of
maxing it out." Python has no security issues with buffers of 64k,
and there is no appreciable memory footprint difference to having a
larger buffer available in uwsgi.
[1]: https://chat.zulip.org/#narrow/stream/31-production-help/topic/works.20in.20Edge.20not.20Chrome/near/719523
[2]: https://nginx.org/en/docs/http/ngx_http_core_module.html#client_header_buffer_size
[3]: https://uwsgi-docs.readthedocs.io/en/latest/ThingsToKnow.html
[4]: https://uwsgi-docs.readthedocs.io/en/latest/Options.html#buffer-size
Support for this header was removed in Chrome 78, Safari 15.4, and
Edge 17. It was never supported in Firefox.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
This check loads Django, and as such must be run as the zulip user.
Repeat the same pattern used elsewhere in nagios, of writing a state
file, which is read by `check_cron_file`.
Replication checks should only run on primary and replicas, not
standalone hosts; while `autovac_freeze` currently only runs on
primary hosts, it functions identically on replicas, and is fine to
run there.
Make `autovac_freeze` run on all `postgresql` hosts, and make
standalone hosts no longer `postgres_primary`, so they do not fail the
replication tests.
These style of checks just look for matching process names using
`check_remote_arg_string`, which dates to 8edbd64bb8. These were
added because the original two (`missedmessage_emails` and
`slow_queries`) did not create consumers, instead polling for events.
Switch these to checking the queue consumer counts that the
`check-rabbitmq-consumers` check is already writing out. Since the
`missedmessage_emails` was _already_ checked via the consumer check, a
duplicate is not added.
Alex Vandiver
Anders Kaseorg
Tim Abbott
Mateusz Mandera
Tran Sang
Matt Keller